[holonsoftFounder]

Phase 3. Every allocation that scaled with field count is gone from both directions. new byte[ no longer appears on either path.

Write side

removed	was
BitConverter.GetBytes	a fresh byte[] per multi-byte field
LINQ .Reverse()	an iterator on top of that array, per big-endian field
List<byte> growth	started empty, doubled its way up; sized from the protocol now
string encoding	a byte[] for the text, a string from char.ToString(), a second byte[] for the fill char — per message

The encoder and encoded fill bytes are resolved once in Prepare(), and text is encoded straight into the list's backing array via CollectionsMarshal.

Read side

removed	was
big-endian scratch buffers	a byte[] per field, built only to feed BitConverter. BinaryPrimitives reads from a span. Ten sites.
Guid	a byte[16], plus a LINQ reverse and another array for big-endian. Now stackalloc, reversed in place
string decoding	copied into a temp byte[] before GetString. Now decoded straight from the source span

The big-endian read allocations were invisible to every test — both byte orders were correct, one was just quietly allocating 3x. Only the memory column surfaced it.

One deliberate subtlety

The fill bytes are produced with the real encoder, not (byte) fillChar. For a character above 0x7F the ASCII encoder substitutes '?' (0x3F) while a raw cast emits the low byte — a silent wire-format change for anyone using a non-ASCII fill character. No golden vector covers that case, so it would have shipped unnoticed.

Against the shipped 3.6.1 code (bcea658)

	before	now	time	alloc
Read LE	118.09 ns	91.79 ns	−22.3%	344 → 104 B
Read BE	146.21 ns	93.33 ns	−36.2%	568 → 104 B (−82%)
Read reused	115.58 ns	94.81 ns	−18.0%	288 → 48 B (−83%)
Write LE	245.33 ns	134.93 ns	−45.0%	888 → 208 B (−77%)
Write BE	376.86 ns	134.06 ns	−64.4%	1528 → 208 B (−86%)
Read w/ string	73.96 ns	64.69 ns	−12.5%	288 → 208 B
Write w/ string	166.05 ns	144.35 ns	−13.1%	384 → 296 B

Byte order is now free in both directions (read 93.3 vs 91.8 ns; write 134.1 vs 134.9 ns).

Golden vectors unmoved throughout — that is the evidence the hand-written endianness arithmetic matches BitConverter byte for byte. 83/83.

🤖 Generated with Claude Code